How Wren AI turns a question into an answer
Ask in plain language. Wren retrieves the business context that matters, writes the SQL, validates it, and runs it against your own database — the data stays where it lives.
Ask a question
A user asks in natural language. No SQL, no need to know table names — just the business question and its context (who's asking, which connected data source).
Why natural language is enough
- Questions resolve against your semantic model (MDL), so “revenue” means what your business defines it to mean — not a raw column guess.
- The same question works from the Wren UI or from any agent that speaks to Wren over MCP.
Retrieve the right context
Instead of dumping the whole schema at the model, Wren AI Service pulls only the fragments that matter: relevant tables and columns, past confirmed queries, and the business rules tied to them.
Handling very large schemas (hundreds–thousands of tables)
- Wren never sends the entire schema. It uses semantic retrieval + schema linking to fetch only relevant fragments.
- When an answer spans several tables, MDL relationships and knowledge rules help complete the joins.
- Prompts stay inside the model's context window regardless of overall schema size.
Generate the SQL
The retrieved context guides the LLM to produce modeled SQL — grounded in your definitions instead of a hopeful guess.
SELECT product_name, SUM(net_revenue) AS revenue
FROM orders
WHERE quarter = current_quarter()
GROUP BY product_name ORDER BY revenue DESC LIMIT 10;
What actually leaves your environment for the model
Sent to the LLM context only
- The user's question
- Schema metadata (tables, columns)
- MDL definitions
- Knowledge rules & semantic context
- Some sampled values (profiling)
- Aggregated results, to summarise the answer
Stays in your environment never shipped
- A bulk copy of your database
- Raw row-level records at large
- Query execution — runs on your source
Validate & execute
Wren AI Core dry-plans and validates the query, transpiles modeled SQL into your database's exact dialect, enforces access control, then executes it against your connected source.
Correctness, not confidently-wrong
- Invalid SQL is caught before it runs and returned as a structured error with hints for a self-correcting retry.
- Transpilation targets 22+ sources (Snowflake, BigQuery, Redshift, Postgres, MySQL…) from one modeled query.
- Execution happens where your data already lives — Wren generates and sends SQL, it doesn't relocate your warehouse.
See the answer
Wren UI shows the result preview, a chart, and smart follow-up questions — and you can save or share it as a GenBI view.
And it gets smarter
- Confirmed question→SQL pairs are stored in memory and recalled on future questions.
- Feedback and reviewed examples seed later retrieval, so repeated question shapes get faster and more accurate.